In most tasks related to opinion mining and sentiment analysis, it is necessary to compute the semantic orientation (i.e., positive or negative evaluative implications) of certain opinion expressions. Recent works suggest that semantic orientation depends on application domains. Moreover, we think that semantic orientation depends on the specific targets (features) that an opinion is applied to. In this paper, we introduce a technique to build domain-specific, feature-level opinion lexicons in a semi-supervised manner: we first induce a lexicon starting from a small set of annotated documents; then, we expand it automatically from a larger set of unannotated documents, using a new graph-based ranking algorithm. Our method was evaluated in three different domains (headphones, hotels and cars), using a corpus of product reviews which opinions were annotated at the feature level. We conclude that our method produces feature-level opinion lexicons with better accuracy and recall that domain-independent opinion lexicons using only a few annotated documents.
展开▼
